CN1374789A - Digital image compression method - Google Patents

Abstract

A method of compressing digital representations of images stores the images in multiple subsampling quality scales in a progressive manner such that a higher quality scale contains only data incremental to the data in an adjacent lower quality scale. The data in each quality scale is transformed, quantized, and entropy encoded. A discrete wavelet cosine transformation combining frequency transformation features of a discrete cosine transformation and spatial transformation features of a wavelet transformation is provided. Sequences of subsampling quality scales are provided for multi-scale representations of images. A novel context selection method is used which treats color components and coefficients of different positions differently. An image compressed in a given quality scale can be transmitted or decompressed progressively allowing progressive downloading or transmission over the Internet.

Description

Digital image compression method

Technical field

The present invention relates to a kind of compression method of digital image, the pictorial data after being characterized in will compressing in a kind of progressive mode is stored according to a plurality of levels of audit quality and is transmitted.

Background technology

The high-quality colour image is carried out stored digital and demonstration becomes more and more general.In order to overcome a large amount of memory requirements and to reduce the transmission time of quality digital image and the cost of spending, data compression method has obtained unprecedented development, and particularly everybody JPEG and the latest edition JPEG2000 thereof that know have become industrial standard.Because compression efficiency and reconstruction quality are a pair of contradiction, so data compression generally all can be weighed between data volume after the compression and image reconstruction quality and be compromised.If image and the original picture rebuild are different, so used data compression method is called " lossy compression method " method.

As you know, in basic JPEG method, image is converted to the form of brightness/chroma usually, is expressed as YUV or YCbCr.Wherein, Y is called main color component or luminance component; U and V or Cb and Cr are called less important color component or chromatic component.The quantity of less important color component can reduce usually, obtains by one group of pixel value is averaged.The data of each component are handled by piecemeal then, each piece carries out discrete cosine transform (DCT), afterwards, all DCT coefficients in each piece are quantized, (being integer promptly) divided by a quantization parameter of reserving in advance and rounding, again according to conditional probability, adopt Huffman encoding or arithmetic coding method to encode to the coefficient after quantizing.A general jpeg file comprises compression parameters, quantization table and entropy coding table in front, so that decoding program can carry out contrary operation.

JPEG also has some optional extended methods except basic skills, wherein a kind of is incremental model, and it can support real-time picture transmission.In incremental model, the DCT coefficient can transmit according to the repeatedly scanning of image being carried out segmentation.Every increase one piece of data, decoder just can produce more visual version of high-quality level.Yet in great majority were realized, the number of pixels that each credit rating is used all was identical.

Although JPEG and JPEG2000 have a very wide range of applications, these two kinds of methods all exist shortcoming.The subject matter of JPEG compression be compression ratio general, blocking artifact and relatively poor progressive transmission quality arranged.In JPEG, in order to reach the desired compression rate, a key operation quantizes the DCT coefficient exactly.Hour compression ratio is lower for quantization parameter, and when quantizing can to bring blocking artifact when coefficient is big, promptly can see the edge of square on the image of rebuilding.Use the JPEG compression method, image quality can not be along with the progressive decline of the raising of compression ratio, and therefore the jpeg picture of a progressive decoding is decoded at its whole levels is to produce satisfied viewing effect before intact.

JPEG2000 is that JPEG2000 has used wavelet transformation for some shortcomings that overcome JPEG design, and it can be along with the fine and smooth more variation of the raising of compression ratio.Yet JPEG2000 has also brought higher computational complexity.In the progressive mode that JPEG2000 uses, be that coding or decoding all need excessive amount of calculation.Though the wavelet transformation among the JPEG2000 has improved along with compression ratio improves and the image quality that degenerates, it does not fundamentally improve the image compression rate, to such an extent as to when image quality was better, its compression ratio and JPEG were roughly suitable.And the arithmetic coding based on the context prediction that JPEG2000 uses does not utilize each this characteristic of chrominance component height correlation in the image.

Therefore, for existing image compression technology, still exist room for improvement.Everybody waits in expectation and a better conversion can occur to reach the compression ratio of execution speed and Geng Gao faster, also to wait in expectation the progressive compression method that has more efficient a, better quality.At last, we also see one can improved place, promptly in the context prediction, utilize the color correlation, and provide a compression method for the color space outside the yuv space.

Summary of the invention

Purpose of the present invention is intended to overcome above-mentioned the deficiencies in the prior art, propose a kind of more efficient, better quality can be in progressive mode with image according to the form of a plurality of credit ratings effectively storage or the compression method of the digital image that transmits by computer network.

Realize the technical scheme of above-mentioned purpose: a kind of compression method of digital image, the numeral of an image is compressed into an one dimension code stream, the numeral of image comprises a two-dimentional pixel array, and wherein each pixel all interrelates with a main colour band and less important colour band, wherein:

Image table is shown as the credit rating that a series of quality is successively decreased gradually, wherein, the data that the better quality grade comprises relatively low-quality level want many, reduce the colour band number of higher quality grade or the number of minimizing higher quality grade pixel and can obtain lower credit rating;

Represent an image according to each credit rating, described image comprises a gross grade image and difference image;

The difference image of some credit ratings is to do difference by two images to obtain, and an image is the realistic images of this grade, and another image is to obtain amplifying at the image reconstruction than low-quality level;

By a process gross grade image and difference image are expressed as integer value, this process comprises: earlier image transform is become the one group of coefficient that interrelates with known function, with quantized value this group coefficient sets is quantized then and rounding arrives integer value;

Coding makes integer-valued numeral produce the code stream of an one dimension corresponding to the integer value of minimum quality grade and difference image with harmless sequencing statistical coding method.

The method of determining image reconstruction comprises: the coefficient that quantized and quantized value are multiplied each other carry out inverse quantization; The inverse transformation of carrying out interrelating with known function produces one and rebuilds expression.

The compression method of digital image of the present invention is mainly used in three-colour image, and image is represented that by main color component and less important chrominance component each color component is all by a two-dimentional pixel array representation.The multicolour space, all at the row of consideration, when being in rgb space, main colour component is meant green as rgb space and YUV brightness/chroma color space, less important colour component is meant redness and blueness; When being in yuv space, main colour component is meant brightness, and less important colour component is meant colourity.

Described credit rating sequence comprises: first credit rating, and wherein each pixel all comprises all colour components; Second credit rating, wherein each pixel all comprises main colour component and a less important colour component; The 3rd credit rating, wherein each pixel all comprises a main colour component, and its quantity is 4 times of each less important colour component quantity; The 4th credit rating, wherein each pixel all only comprises a colour component, and the quantity of main colour component is the twice of less important colour component quantity; The 5th credit rating is derived from first credit rating, by both direction all divided by an integer ratio factor, reduce the number of pixels of its horizontal direction and vertical direction; The 6th credit rating is derived from the 4th credit rating, and wherein each pixel comprises main colour component and a less important colour component; The 7th credit rating derives from the 4th credit rating, and wherein each pixel all comprises a main colour component, and its quantity is 4 times of each less important colour component quantity.

Gross grade image and difference image are converted to one group of coefficient that interrelates with known function.In typical implementation procedure, minimum quality grade image and difference image before conversion by piecemeal.In traditional JPEG method, employing be discrete cosine transform.Among the present invention, the transform method that adopts is discrete wavelet cosine transform (DWCT), the two-dimensional transform that this conversion combines many resolutions feature of the frequency domain characteristics of discrete cosine transform and spatial transform (as haar wavelet transform), the process of two-dimensional transform comprises: before doing conversion, the data of minimum quality grade and difference image are resolved into the square of M * M, and M is 2 integer power; If the piece that size is J * K, J and K are not 2 integer powers, at first will mend into it the square of a M * M, can realize so that after the conversion, J * K nonzero value being arranged at the most in the coefficient sets by replenish new element in the original block of J * K; Two-dimensional transform comprises the one-dimensional transform of two-dimensional array being done a line direction, again this array is done an one-dimensional transform on the column direction; Be based on the one-dimensional transform of using discrete cosine transform and permutation function adopt the recursive fashion definition, and the output element of conversion can be divided into even number element and odd elements, and the even number element comprises the expression than low-quality level in the input data of conversion.DWCT is not only faster than traditional wavelet transformation speed, and compares with the conversion of using in the past, can produce compacter data and distribute.The DWCT coefficient uses the value of appointment in the quantization table to quantize, and rounding is to integer.

The quantization parameter of gross grade and difference image compresses with a kind of harmless arrangement statistical coding method, and this arrangement statistical coding method comprises context prediction, ordering and the entropy coding of quantization parameter, wherein:

In the context forecasting process, the value of each coefficient is predicted by the adjacent image point coefficient value that each color component is predicted respectively: for main colour component, context comprises the adjacent coefficient of a location index and main color pixel; For first less important colour component, context comprises adjacent coefficient in a location index, this colour band and the coefficient that is positioned at the main colour band of same position; For second less important colour component, context comprises a location index, the adjacent coefficient in this colour band and be positioned at main colour band coefficient and first less important colour band coefficient on the same position.

In above-mentioned context Forecasting Methodology, quantization parameter is divided into four groups: first group of zero coefficient that comprises corresponding to the extreme lower position index according to the position in array; Second group comprises all data of removing the row of first outside first group of data; The 3rd group comprises all first columns certificates of removing outside first group of data; The 4th group comprises all residual coefficients; The context of each group is all different.

Sorting operation is that two-dimensional array is arranged in one-dimension array, in order to make the correlation between data reach maximum, used a kind of quaternary tree sort method, in this quaternary tree sort method, two dimension coefficient array is divided into 4 equal zones, is respectively upper left, upper right, lower-left and bottom right, and that each zone is divided into again is upper left, upper right, the subregion of lower-left and four the identical sizes in bottom right, this process constantly repeats, till each subregion all only comprises a pixel.Quantize or the context prediction before to carry out ordering, and all related datas coefficient for example, quantization table and contextual mapping relations also will be retained.

At last, again this data value through context prediction and ordering is carried out entropy coding (as arithmetic coding), the entropy coding process is that the numeral of image is compressed into a code stream, numeral comprises the pixel array of a two dimension, wherein each pixel all comprises a main colour component and less important colour component, and the probability tables that will use in the entropy coding process is formed by the context Forecasting Methodology.

The decompression method of above-mentioned digital image comprises the steps:

Recover the context prediction probability table of use in the compression;

Code stream is decoded, recover integer value corresponding to gross grade image and difference image;

The adjustment order is arranged in a two-dimensional array form to decoded integer;

Each desorption coefficient corresponding to gross grade and difference image is multiplied each other with quantized value;

Do the inverse transformation that interrelates with known function and rebuild the numeral of gross grade and difference image;

The image that is positioned at than low-quality level is extended to next higher quality grade;

A given credit rating, difference image and the visual addition that is extended to this grade are obtained the numeral of a reconstruction of this grade.

Decode procedure is the inverse process of coding substantially, produces image reconstruction.

In the decompression method of above-mentioned digital image, with inverse transformation that known function interrelates is the conversion that frequency domain transform characteristics and many resolutions transform characteristics are combined, this conversion is to carry out recursive definition on the basis of an inverse discrete cosine transformation and a permutation function, wherein, the input data of inverse transformation have been divided into even number element and odd elements, and the even number element comprises the more low-level expression in the input data of this conversion.

Credit rating realizes that by the sub-sampling sequence in a sample sequence, the length of sampled data is successively decreased, sampled data represents to comprise the pixel array of a two dimension, wherein, each pixel all has main colour component and less important colour component, and this sequence comprises following particular sequence:

First kind of sampled representation, each pixel all has all colour components;

Second kind of sampled representation, each pixel all have main colour component and a less important colour component;

The third sampled representation, each pixel all have a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity;

The 4th kind of sampled representation comes from first kind of sampled representation, but the dimension of its two-dimensional array is divided by an integer ratio factor, the corresponding in the horizontal and vertical directions number that reduces pixel;

The 5th kind of method of sampling is derived from the 4th kind of sampled representation, and each pixel all comprises main colour component and a less important colour component;

The 6th kind of sampled representation is derived from the 4th kind of sampled representation, and each pixel all comprises a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity.

Credit rating also can realize by another kind of sub-sampling sequence, in this sample sequence, the length of sampled data is successively decreased, sampled data represents to comprise the pixel array of a two dimension, wherein, each pixel all has main colour component and less important colour component, and this sequence comprises following particular sequence:

First kind of sampled representation, each pixel all has all colour components;

Second kind of sampled representation, each pixel all have main colour component and a less important colour component;

The third sampled representation, each pixel all have a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity;

The 4th kind of sampled representation, each pixel comprises a color component, and the quantity of main color component is the twice of less important color component quantity.

Adopt technique scheme, the technological progress that the present invention gives prominence to is: 1, image table is shown as the credit rating that a series of quality is successively decreased gradually, adopting the better quality grade to use represents the adjacent difference than low-quality level, each difference image only comprises corresponding to the data increment than low-quality level, thereby, can be in progressive mode with the form storage of image according to a plurality of sub-sampling credit ratings.Image according to the present invention's compression can transmit by computer network or internet, thereby realizes progressive watching and downloading.More have, a browser can be watched an images in the credit rating of certain appointment, and to the data that are higher than this credit rating ignore no matter.2, difference image is that the basis obtains with the image reconstruction of adjacent lower grade, has avoided the accumulation of error, and image quality is higher.3, adopt discrete wavelet cosine transform (DWCT), the frequency domain characteristics of discrete cosine transform and many resolutions feature of spatial transform (as haar wavelet transform) are combined, DWCT is not only faster than traditional wavelet transformation speed, and compare with the conversion of using in the past, can produce compacter data distribution.4, because progressive, the continuous variation universal law that is a width of cloth picture, image elements forward and backward, about have extremely strong correlation.The present invention catches this principal character, in the context forecasting process, the value of each coefficient predicted with each color component by the adjacent image point coefficient value predict respectively, and adopt the quaternary tree sort method in sorting operation, the correlation between data improves greatly.In view of the above, the compression ratio of pictorial data can improve greatly, can save memory space or transfer resource fully, effectively.

Description of drawings

Fig. 1 is a flow chart that still image is compressed according to the technology of the present invention;

Fig. 2 separates compressed bit stream a flow chart that is pressed into digital image according to the technology of the present invention;

Fig. 3 is four group categories that are used for the context prediction according to the technology of the present invention.

Embodiment

Below in conjunction with accompanying drawing, the present invention is further detailed explanation by embodiment.

The method of the compressed digital image that proposes among the present invention is shown as a plurality of sub-sampling credit ratings to image table in a kind of progressive mode, does the accumulation that can avoid error like this.In this application, the method that still image is dealt with only has been discussed.Because no matter the every frame image in which kind of video sequence can be regarded still image as, so this method also can be used video image.The image that this method is handled is mainly three-colour image: a main color component (representing with P) and less important color component (being expressed as S and Q).Certainly, this method equally also can expand to monochrome image or multicolor image deals with.P, S and Q component can be distinguished G (green), R (redness) and B (blueness) component of corresponding rgb color space, also can distinguish Y, U and the V component of corresponding YUV brightness/chroma color space.In typical case, the input of this data compression method is a two-dimentional pixel array, and wherein each pixel all comprises three kinds of color component elements, and output is the one dimension code stream of packed data.

The compression method that proposes among the present invention can run on general calculation machine processing unit (CPU) or digital signal processor (DSP) with form of software, also can realize on the VLSI chip with example, in hardware, or realize in the combining environmental of soft, hardware.When in software, realizing, be used for realizing that the computer instruction of compression method is stored in CPU or the internal memory of DSP, like this, herein device one speech of Shi Yonging or refer to have prelist the instruction dedicated hardware device, or refer to all-purpose computer and have the internal memory of store instruction, or refer to the coupling apparatus of specialized hardware and computer.Image after the compression can be kept in the internal memory, or is provided with demonstration is recovered and done in the back on monitor usefulness, or for the usefulness of doing transmission in in-house network and extranets such as internet.

Fig. 1 has provided use that the inventive method compresses pictorial data one roughly flow process.Be decomposed into a plurality of credit ratings at sub-sampling stage 110, one original pictures, from the highest credit rating, till the minimum quality grade.The first water grade is also referred to as compression quality grade (CQS), and the minimum quality grade is that the gross grade is also referred to as sub-sampling credit rating (SQS).The pictorial data of gross grade SQS is exported at first, then is the data of next credit rating, till the data of CQS are output.

Difference block 120 has provided the sub-sampling image of difference form, and is previous than the data that increase on the low-quality level basis so that each difference image all only is included in.Difference block comprises module 125 to 170, be used for determining poor between a sub-sampling image and the reference picture, at first at the contiguous sub-sampling image that obtains a reconstruction on than low-quality level, then this image reconstruction is amplified and obtain a reference picture, then try to achieve poor between current sub-sampling image and the reference picture.As to be described in detail below, image reconstruction is to obtain by following flow process, at first the sub-sampling image is carried out conversion (130) and quantization operation (140), these two processes can cause damage, and then recover image by inverse quantization module (150) and inverse transform block (160).The output of difference block 120 be the gross level data after quantizing and quantize after difference image.

Basic image after the quantification and difference image are changed into compressed bit stream by harmless sequencing statistical coding module.The sequencing statistical coding module comprises 185, one order module 190 of a context prediction module (it changes into one-dimension array with two-dimensional array), also has a harmless entropy coding module 195, as arithmetic coding.Decompressing method is the inverse process of compression process as shown in Figure 2 basically, below also can be detailed discuss.

In specific implementation process, sub-sampling module 110 has defined 7 sub-sampling grades (also claiming 7 kinds of sampled representation) altogether, and different sub-sampling sequences all have similar definition.These 7 sub-sampling grades are made as the highest credit rating of 0 to 6,0 representative, and 6 is the minimum quality grade.Each grade numeral is carried out a kind of specific sub-sampling operation to pictorial data.

0 grade of sub-sampling represents that (also claiming first kind of sampled representation) is 4: 4: 4, and promptly original picture does not just carry out sub-sampling at all.All pixels all remain with P, S, Q component element, and are as shown in table 1.

Table 1:0 level sub-sampling 4: 4: 4

??(P，S，Q)	??(P，S，Q)	??(P，S，Q)	??(P，S，Q)
				??(P，S，Q)	??(P，S，Q)	??(P，S，Q)	??(P，S，Q)
??(P，S，Q)	??(P，S，Q)	??(P，S，Q)	??(P，S，Q)
				??(P，S，Q)	??(P，S，Q)	??(P，S，Q)	??(P，S，Q)

1 grade of sub-sampling represents that (also claiming second kind of sampled representation) is 4: 2: 2, and fundamental component P does not carry out sub-sampling, has only S and Q component by sub-sampling.Have 6 kinds of patterns in this sampling grade, shown in following table 2-7.

4: 2: 2 patterns 1 of table 2:1 level sub-sampling

??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)
				??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)
??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)
				??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)

First kind of pattern in 4: 2: 2 sub-sampling modes is identical with the sub-sampling method of using in television standard (as NTSC and PAL) and motion video transmission standard MPEG2.In table, " x " represents a component elements of removing.In second kind of pattern as shown in table 3, S and Q component are the diagonal angle to be arranged, so that in S component and the Q component, along continuous straight runs still vertically interval all is uniform.Pattern 2 is sub-sampling patterns first-selected in 1 grade of sampling.

4: 2: 2 sub-sampling patterns 2 of table 3:1 level sampling

????(P，S，x)	????(P，x，Q)	????(P，S，x)	????(P，x，Q)
				????(P，x，Q)	????(P，S，x)	????(P，x，Q)	????(P，S，x)
????(P，S，x)	????(P，x，Q)	????(P，S，x)	????(P，x，Q)
				????(P，x，Q)	????(P，S，x)	????(P，x，Q)	????(P，S，x)

Mode 3 has a little different with pattern 2, and promptly the position of S and Q exchanges.As shown in table 4.

4: 2: 2 sub-sampling mode 3s of table 4:1 level sampling

??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)
				??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)
??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)
				??(P，S，x)	??(P，x，Q)	??(P，S，x)	??(P，x，Q)

The 4th kind of pattern is the transposition of pattern 1, and promptly horizontal direction at interval.

4: 2: 2 sub-sampling patterns 4 of table 5:1 level sampling

????(P，S，x)	????(P，S，x)	????(P，S，x)	????(P，S，x)
				????(P，x，Q)	????(P，x，Q)	????(P，x，Q)	????(P，x，Q)
????(P，S，x)	????(P，S，x)	????(P，S，x)	????(P，S，x)
				????(P，x，Q)	????(P，x，Q)	????(P，x，Q)	????(P，x，Q)

Pattern 5, pattern 6 and pattern 1, pattern 4 have a little different, and promptly exchange has taken place in the position of S and Q.

4: 2: 2 sub-sampling patterns 5 of table 6:1 level sampling

??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)
				??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)
??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)
				??(P，x，Q)	??(P，S，x)	??(P，x，Q)	??(P，S，x)

4: 2: 2 sub-sampling patterns 6 of table 7:1 level sampling

??(P，x，Q)	??(P，x，Q)	??(P，x，Q)	??(P，x，Q)
				??(P，S，x)	??(P，S，x)	??(P，S，x)	??(P，S，x)
??(P，x，Q)	??(P，x，Q)	??(P，x，Q)	??(P，x，Q)
				??(P，S，x)	??(P，S，x)	??(P，S，x)	??(P，x，Q)

The secondary sub-sampling represents that (also claiming the third sampled representation) is 4: 2: 0, and fundamental component P still all keeps, and only S and Q is carried out sub-sampling.By S and Q being sampled, also can produce many kinds of patterns in different positions.This paper only enumerates wherein 5 kinds of patterns the most useful.

4: 2: 0 sub-sampling patterns 1 of table 8:2 level sampling

4: 2: 0 forms of the YUV that uses in pattern 1 and MPEG2 and the Joint Photographic Experts Group are similar.In this pattern, per 4 pixels that contain the P element are shared a pair of S and Q element, and they are positioned at the center of 4 P pixels.In second kind of pattern, S and Q component do not overlap.They are the diagonal angle arranges, and is had by different P pixels respectively, as shown in table 9:

4: 2: 0 sub-sampling patterns 2 of table 9:2 level sampling

??(P，S)	????P	????(P，S)	????P
				????P	????(P，Q)	????P	????(P，Q)
????(P，S)	????P	????(P，S)	????P
				????P	????(P，Q)	????P	????(P，Q)

Other three kinds of variants that pattern all is a pattern 2, only S is distributed in different positions with Q.As show shown in the 10-12.

4: 2: 0 sub-sampling mode 3s of table 10:2 level sampling

????(P，Q)	????P	????(P，Q)	????P
				????P	????(P，S)	????P	????(P，S)
????(P，Q)	????P	????(P，Q)	????P
				????P	????(P，S)	????P	????(P，S)

4: 2: 0 sub-sampling patterns 4 of table 11:2 level sampling

????P	????(P，Q)	????P	????(P，Q)
				????(P，S)	????P	????(P，S)	????P
????P	????(P，Q)	????P	????(P，Q)
				????(P，S)	????P	????(P，S)	????P

4: 2: 0 sub-sampling patterns 5 of table 12:2 level sampling

????P	????(P，S)	????P	????(P，S)
				????(P，Q)	????P	????(P，Q)	????P
????P	????(P，S)	????P	????(P，S)
				????(P，Q)	????P	????(P，Q)	????P

3 grades of sub-samplings (also claiming the 4th kind of sampled representation) are called the pattra leaves pattern, and it is a kind of common sample format in the image sensor technology.It comprises 4 kinds of patterns, and each pattern has all defined a kind of specific chrominance component structure.As shown in table 13, in first kind of pattern, each pixel only comprises a color component element.The number of main color component element is the twice of less important color component element number.

Table 13:3 level sampling pattra leaves sub-sampling pattern 1

????P	????S	????P	????S
				????Q	????P	????Q	????P
????P	????S	????P	????S
				????Q	????P	????Q	????P

Other form of pattra leaves pattern is only reset the position of component.Shown in following table 14-16.

Table 14:3 level sampling pattra leaves sub-sampling pattern 2

????P	????Q	????P	????Q
				????S	????P	????S	????P
????P	????Q	????P	????Q
				????S	????P	????S	????P

Table 15:3 level sampling pattra leaves sub-sampling mode 3

????S	????P	????S	????P
				????P	????Q	????P	????Q
????S	????P	????S	????P
				????P	????Q	????P	????Q

Table 16:3 level sampling pattra leaves sub-sampling pattern 4

????Q	????P	????Q	????P
				????P	????S	????P	????S
????Q	????P	????Q	????P
				????P	????S	????P	????S

In the level Four sub-sampling (also claiming the 5th kind of sampled representation), the number of pixel has been reduced.If the size that each pixel shows all keeps identical, then in the level Four sampling, the whole dimension that image shows is with less; On the contrary, if the dimension of picture that shows does not reduce, each pixel is described so, and region occupied will be bigger in image.Color component in pattern is represented in the zero level sampling at 4: 4: 4.In the level Four sub-sampling, an original picture I with M * N pixel _{M * N}Dwindle into size with vertical direction in the horizontal direction at first respectively and be the visual I of m * n _{M * n}, wherein

m＝(M+L-1)/L，n＝(N+L-1)/L.

Then, use 4: 4: 4 forms to represent I _{M * n}In most cases, L is 2 or 3.

Many algorithms all proportionally factor L to visual I _{M * N}Dwindle respectively with vertical direction in the horizontal direction, comprise simple sampling, bilinearity convergent-divergent and two cubes of convergent-divergents, not only be confined to these algorithms certainly.Simple sampling pantography keeps a pixel according to even interval in each L * L piece, and other convergent-divergent rule is to calculate the pixel block that a pixel value is represented L * L according to the neighborhood pixel.The bilinearity pantography is inserted pixel value according to a linear function of level, vertical coordinate at a reposition.The coefficient of function or reserve is in advance perhaps estimated according to the known pixel value of neighborhood.Same, in two cubes of pantographys, interpolating function is the cubic function of horizontal coordinate and vertical coordinate.

In the Pyatyi sub-sampling (also claiming the 6th kind of sampled representation), a size is the visual I of M * N _{M * N}Dwindle into the visual I that a size is m * n according to a scale factor respectively in level and vertical direction _{M * n}, and carry out sub-sampling by 4: 2: 2 forms.Same, in six grades of sub-samplings, image dwindles respectively with vertical direction in the horizontal direction according to scale factor L, and carries out sub-sampling according to 4: 2: 0 forms.In the multiple dimensioned expression of a compressed image, same scale factor L for example equals 2 or 3, will be used for level Four to six grade sub-sampling mode, and therefore, four to six grades of sample modes can obtain same number of pixels.

Dwindle with interpolator arithmetic and often use in pairs, be expressed as (Lx ↓, Lx ↑), wherein L is a scale factor, Lx has then defined a convergent-divergent algorithm.The selection convergent-divergent algorithm how to optimize depends on the feature of pictorial data.Because be to occur in pairs, therefore same L both had been used to dwindle algorithm (Lx ↓), was used for interpolator arithmetic (Lx ↑) again.Some colour band or whole colour band to an images carry out zoom operations, and be general based on above-mentioned sub-sampling pattern.The decision of sub-sampling pattern is carried out sub-sampling with which kind of sub-sampling grade to this images.

In order to form a sub-sampling image sequence, at first, original picture I must belong to a kind of in zero level described above, one-level, secondary and the three grades of sub-sampling ranks.To the image of 4: 4: 4 patterns proportionally the factor 2 only along continuous straight runs UV colour band or RB colour band are dwindled, just obtained an image of 4: 2: 2 patterns.The reduction operation of this along continuous straight runs is carried out by 2x ↓ algorithm.For 4: 4: 4 patterns and 4: 2: 2 patterns, its YUV colour band is the same with the RGB colour band in the size of vertical direction.Equally, to the image of 4: 2: 2 patterns proportionally the factor 2 only vertically to the UV colour band or the RB colour band dwindles the image that can obtain 4: 2: 0 patterns.To the P components image of the image of 4: 2: 0 pattern in the horizontal direction or vertical direction do to dwindle and can obtain the pattra leaves format pattern, in this case, can not do reduction operation simultaneously at both direction.

If being the original scale of visual I, G represents.I can regard the sub-sampling image that obtains as when the zero level sub-sampling, be expressed as I _GFor I _G, we adopt Lx ↓ algorithm to carry out sub-sampling and obtain I _G+1For I _G+1, we adopt algorithm Lx ↓ carry out sub-sampling to obtain I once more _G+2, and the like, arrived sub-sampling credit rating S until us.We have obtained a sub-sampling image sequence I according to the method _G: { I _G, I _G+1... I _S.Obviously, SQS value S can not be littler than original visual grade G.

For example, suppose G=2, S=4.I ₂Be 4: 2: 0 format patterns.So from I _G, we use simple sampling algorithm I ₂Be lowered into a kind of pattra leaves pattern I ₃, shown in table 13-16.We also use given algorithm 2x ↓ with I ₂The P component dwindle, I ₂Be reduced to I ₄, formed 4: 4: 4 format patterns that size is less.Certainly, also can be by using given algorithm 2x ↓ to I ₀Institute's chromatic colour colour band do sampling and directly obtain I ₄

Each chrominance component in the pattra leaves form is not on same position.When a kind of pattra leaves format pattern sub-sampling is become other pixel format of a kind of even lower level, need think better of the arrangement of pixel.Therefore, ensuing two kinds of sample sequences are particularly useful:

Sequence I:{0,1,2,4,5,6}

Sequence II:{0,1,2,3}

For the SQS that is worth less than 3, sequence I and sequence II can use.For the value that the SQS value equals 3, should use sequence II.Greater than 3 value, should use sequence I for the SQS value.In sequence II, the first-selected simple sampling pantography of one-level sampling and secondary sampling.In sequence I, 3 grades of samplings will be skipped in order to avoid produce position vacancy problem.Can use the more complicated method of sampling for sequence I.

For sequence I and sequence II, the difference image sequence has different definition.Under the situation of sequence II, the sub-sampling sequence comprises { I ₀, I ₁, I ₂, I ₃A subclass.When using simple sampling pantography to carry out sub-sampling, I _S+1Clearly be I _SA subclass, S=0,1,2.I _SAnd I _S+1The visual D of difference _SMay be defined as and be contained in I _SIn but be not contained in I _S+1In unnecessary pixel or colour band element.Thereby, D _SCan be written as

D _k＝I _k-I _k+1，k＝0，1，2，

For simplicity, make D ₃=I ₃Like this, we just can obtain a difference sub-sampling image sequence:

{D ₀，D ₁，D ₂，D ₃}.

Under the situation of sequence I, the sub-sampling sequence comprises { I ₀, I ₁, I ₂, I ₄, I ₅, I ₆A subclass.In order to obtain difference image, we with image from being amplified to higher quality level than low quality level.For example, we use given algorithm 2x ↑ amplify I ₄Thereby the P component obtain the visual J of a higher level ₂..Equally, we use given algorithm 2x ↑ with I _kS component and Q component amplify to obtain higher level image J _K-1, k=6,5,2,1.The difference image of rank k may be defined as then:

D _k＝I _k-J _k，k＝0，1，2，4，5.

Make D ₆=I ₆, can provide difference sampling image sequence:

{D ₀，D ₁，D ₂，D ₄，D ₅，D ₆}.

Difference method among the sequence I also can be applicable among the sequence II.

According to the technology of the present invention, except above-mentioned difference sampling image sequence, the difference image that produces in 120 modules in Fig. 1 has also adopted another form.Above-described difference table is shown with a shortcoming: propagated error promptly can the process from the minimum quality grade to the better quality grade.In order to avoid the accumulation of error in current difference step, the reference picture that is used for subtracting each other is replaced by image reconstruction, and image reconstruction is recovered by the data than low quality level.

We only consider the situation of sequence I here.The situation of sequence II is simpler, can do similar processing to it.A given SQS S and a CQS C at first obtain its sample sequence { I _C, I _C+1..., I _S.Gross grade image I _STo do through conversion module 130 and quantization modules 140 and diminish processing.Conversion module 130 can resemble and use discrete cosine transform or wavelet transformation respectively JPEG or the JPEG2000 method.In this application, conversion module 130 can use a new discrete cosine wavelet transformation, and specific definition is arranged below.

By the above-mentioned process that diminishes is carried out inverse operation and is come image reconstruction I _S, we can obtain G _SG generally speaking _SAnd I _SBe different, because module 130 and module 140 all diminish.Use a given algorithm Lx ↑, we can be with G _SBe amplified on the higher quality level (S-1), and obtain H _S-1We can obtain D from following formula then _S-1:

D _S-1＝I _S-1-H _S-1?????????????????????????????(1)

D _S-1Also be quantized through conversion.Make D ' _S-1Expression is through conversion and the D through quantizing _S-1

Executive Module 130 and 140 reverse process are rebuild D ' _S-1, we obtain G _S-1Equally, G _S-1And D _S-1Also different.Based on G _S-1And H _S-1, we can rebuild I from following formula _S-1:

F _S-1＝G _S-1+H _S-1?????????????????????????????(2)

In the following formula, F _S-1Be to original picture I on (S-1) rank _S-1One approximate.Note F _S-1And I _S-1Between difference and D _S-1And G _S-1Between difference identical, can both prevent the mistake transmission that begins from the S rank.Now, use given 2x ↑ algorithm, we are F _S-1Be amplified on the higher level (S-1), and obtain H _S-2We obtain D from following formula then _S-2:

D _S-2＝I _S-2-H _S-2

Now to D _S-2Do conversion and quantification obtains D ' _S-2, by D ' _S-2Obtain G _S-2This process continues until the difference image D at first water grade C always _CTill being determined, being transformed, being quantized.Image H _kWith J defined above _kDifferent places are: H _kBe to rebuild out with the data that diminish after more low-level conversion and the quantification, and J _kBe to use than low-quality level image I _K+1Amplification obtains.

Computational methods described above in Fig. 1 by flowcharting.Difference between the enlarged image of subtraction block 125 calculating sub-sampling images and reconstruction, the latter is clearly provided by (S-1) grade in the equation (1).The output D ' of quantization modules 140 had both passed to sequencing statistical coding module 180, passed to inverse quantization module 150 again to obtain the difference image.The output of inverse transform block 160 is numerical value G, and it is the reconstruction to D '.As in equation (2), under grade (S-1), addition module 165 amplifies the visual H addition that forms the visual G of difference that rebuilds with by adjacent lower grade.S has for the gross grade: D _S=I _SAnd H _S=0, flow chart description whole differential process, produced in this process a image sequence through transform and quantization D ' _S, D ' _S-1... D ' _C.

Conversion module 130 uses canonical transformation such as discrete cosine transform (DCT), wavelet transformation (WT) or other conversion to obtain conversion coefficient.In the present invention, a kind of conversion of novelty: discrete cosine wavelet transformation (DWCT) is used in the conversion module.The DWCT conversion combines the spatial domain characteristics of Ha Er (small echo) conversion and the frequency domain characteristics of DCT.

Below in the formula, bold-type letter for example X or Y is used for representing a vector.Footnote is used for representing the dimension of vector.X _NRepresent the vector that N element arranged, its n element representation is X _N[n].Round parentheses () are used for comprising the variable (as f (x)) of a function, and square brackets [] are used for comprising the sequence number (as X[n]) of a vector element.We represent a kind of linear transformation with F (X), and it also can be expressed as a matrix multiplication FX, and wherein F is the matrix form of conversion F.

If X _N=[x ₀, x ₁..., x _N-1] ^TAnd Y _N=[y ₀, y ₁..., y _N-1] ^TBe two vectors, insert function so: Interleave (X with N element _N, Y _N) can produce a vector with 2N element:

Interleave(X _N，Y _N)＝[x ₀，y ₀，x ₁，y ₁，…，x _N-1，y _N-1] ^T.

A given vector X _N=[x ₀, x ₁..., x _N-1] ^T, the swap block T that size is N _N(X _N) definition of following matrix form arranged:

T _N(X _N)＝T _NX _N，

T _N(X _N)＝T _N([x ₀，x ₁，…，x _N-1] ^T)＝[x ₀，x ₂，…，x _N-2，x _N-1，…，x ₃，x ₁] ^T.

Wherein, T _NBe certain matrix, T _N(X _N) be reversible.

DWCT is a recursive definition, for the contact between DWCT and DCT that uses previously and the wavelet transformation is described, at first provides the recursive definition of a DCT.For one 2 n dimensional vector n X ₂, dct transform C ₂(X ₂) be defined as:

N n dimensional vector n X _NDCT, wherein N is 2 integer power, is expressed as:

Y _N＝C _N(X _N)＝C _NX _N，

Herein, C _NDefined an one dimension DCT who is of a size of N, E _N/2And O _N/2Be defined as:

E _N/2[n]=Y _N[2n] O _N/2[n]=Y _N[2n+1], n=0,1 ..., N/2-1. and

Y _N＝Interleave(E _N/2，O _N/2).

Dimension is that the DCT of N can be that the DCT form of N/2 is come recursive definition based on dimension:

E _N/2＝C _N/2(U _N/2)＝C _N/2U _N/2，

Q _N/2＝C _N/2(V _N/2)＝C _N/2(V _N/2)，

O _N/2[n]=Q _N/2[n]+Q _N/2[n+1], n=0,1 ..., N/2-1 wherein $U_{N/2}\left[n\right]=\left(X_N\right[n]+X_N[N-n-1\left]\right)/\sqrt{2},n=\mathrm{0,1},...,N/2-1,$ $V_{N/2}\left[n\right]=\left(X_N\right[n]-X_n[N-n-1\left]\right)/\left(2\sqrt{2}\cos \right[(2n+1)\mathrm{\π}/2N\left]\right),n=\mathrm{0,1},...,N/2-1,$

Note Q _N/2It is a vector that has only N/2 element.In order to simplify computing, Q _N/2Can extend to N/2+1 element, and Q _N/2[N/2]=0.Clearly, E _N/2[n], Q _N/2[n] and C _NAll the DCTs by low dimension N/2 defines.

The DCT that obtains by above-mentioned recursive definition examines C _NIt or not standard.In order to obtain the DCT nuclear of standard, as long as the DCT of the standard onrecurrent definition of promptly given dimension N is with C _NFirst the row multiply by

Get final product.

Haar transform (HT) also can recursive definition.For a two-dimensional vector X ₂, H ₂(X ₂) be the same with DCT:

A N n dimensional vector n X _NHT be expressed as:

Y _N＝H _N(X _N)＝H _NX _N，

H herein _NDefined an one dimension HT who is of a size of N, E _N/2And O _N/2Be defined as:

E _N/2[n]=Y _N[2n] O _N/2[n]=Y _N[2n+1], n=0,1 ..., N/2-1. wherein

Y _N＝Interleave(E _N/2，O _N/2).

The HT that is of a size of N can be based on the HTs recursive definition that is of a size of N/2:

E _N/2＝H _N/2(U _N/2)＝H _N/2U _N/2，

O _N/2=V _N/2Wherein $U_{N/2}\left[n\right]=\left(X_N\right[2n]+X_N[2n+1\left]\right)/\sqrt{2},n=\mathrm{0,1},...,N/2-1,$ $V_{N/2}\left[n\right]=\left(X_N\right[2n]-X_N[2n+1\left]\right)/\left(\sqrt{2}\right),n=\mathrm{0,1},...,N/2-1,$

Clearly, E _N/2And H _NHTs definition by lower-order N/2.

The same with the recursive definition of DCT, the nuclear H of the HT that obtains above _NNeither standard, with H _NFirst the row be multiplied by

Promptly obtain the standard HT of given dimension N.

The one dimension DWCT that the present invention proposes also is a recursive definition.At first, be of a size of 2 DWCT W ₂(X ₂) definition identical with DCT and HT: $Y_2=W_2\left(X_2\right)=C_2\left(X_2\right)=H_2\left(X_2\right)=W_2{X}_2=\left[\begin{array}{cc}1& 1\\ \frac{1}{\sqrt{2}}& -\frac{1}{\sqrt{2}}\end{array}\right]X_2.....\left(3\right)$

Next step, the DWCT that is of a size of N may be defined as: (N is 2 integer power)

Y _N＝W _N(X _N)＝W _NX _N?????????????????????????(4)

Two vector E _N/2And O _N/2Be defined as:

E _N/2[n]=Y _N[n], O _N/2[n]=Y _N[N+n], n=0,1 ..., N/2-1. wherein $Y_N=\left[\begin{array}{c}{E}_{N/2}\\ O_{N/2}\end{array}\right]...........\left(5\right)$

Although we for convenience and DCT, HT comparison, still E _N/2And O _N/2Be called even number part and odd number part, in fact in DWCT, the even number part is partly separated with odd number, does not intersect and deposits.

Recursive definition on the DWCT that is of a size of N now can go out at the base of the DWCTs that is of a size of N/2:

E _N/2＝W _N/2(U _N/2)＝W _N/2U _N/2???????????????????????(6)

Q _N/2＝C _N/2(V _N/2)＝C _N/2V _N/2???????????????????????(7)

O _N/2[n]=Q _N/2[n]+Q _N/2[n+1] n=0,1 ..., N/2-1 (8) wherein $U_{N/2}\left[n\right]=\left(Z_N\right[n]+Z_N[N-n-1\left]\right)/\sqrt{2}$ $=\left(X_N\right[2n]+X_N[2n+1\left]\right)/\sqrt{2},n=\mathrm{0,1},...,N/2-1,.....\left(9\right)$ $V_{N/2}\left[n\right]=\left(Z_N\right[n]-Z_N[N-n-1\left]\right)/\left(2\sqrt{2}\cos \right[(2n+1)\mathrm{\π}/2N\left]\right)$ $=\left(X_N\right[2n]-X_N[2n+1\left]\right)/\left(2\sqrt{2}\cos \right[(2n+1)\mathrm{\π}/2N\left]\right)n=\mathrm{0,1},...,N/2-1,...\left(10\right)$ Z _N＝T _N(X _N)＝T _N(X _N)

Same, we still suppose Q _N/2[N/2]=0.

In the recursive definition of DCT and HT, the nuclear W of Ding Yi DWCT herein _NBe do not have standardized.At W _NFirst trip be multiplied by

Can obtain the DWCT nuclear of a standard.Certainly, the DWCT conversion is not necessarily to need standardization.In further discussing below, need consider to the selection of the quantized value of use in the quantization modules 150 whether DWCT is standard.

Though W _N(X _N) it seems point image compound function C _N(T _N(X _N)), but be actually distinguishing.T _N(X _N) only change X when being of a size of N _NData sequence, and W _N(X _N) in recursive operation each time, all change the order of data.From another angle, define on this meaning based on DCT and a permutation function from DWCT, DWCT can be regarded as the distortion of DCT.Permutation function can not change the value of input, can only change its order.

The discrete wavelet cosine transform is gained the name because of equation (6), (7) and (9).The even number part E of the DWCT of definition in equation (6) and (9) _N/2With the even number of Haar transform partly is identical.Odd number part of O in the equation (7) _N/2Be based on DCT definition.Therefore the DWCT place different with HT only is the odd number part, and the odd number part is corresponding to high-frequency information.Vector Y _NElement be used as the DWCT coefficient.In fact, the DWCT coefficient can come from the HT coefficient, is some additional calculations that have in its odd number part correspondence.Therefore DWCT also is a kind of multi-resolution representation of data, and the even number of data partly comprises representing than low quality level of input data.Ordinary meaning, DWCT are a kind of wavelet transformations because each time in the recurrence DWCT all be taken as a Haar transform through operation sequential.Final result is the expression of a compactness, and wherein other coefficient of lowermost level has the most information of image.

Here Ding Yi DWCT is that DWCT can use in any data compression method at the difference image that comes from data reconstruction.For example, in different data compression methods, as in JPEG or JPEG2000, DCT or wavelet transformation replace with DWCT respectively.

Obviously, W _N(X _N) inverse transformation exist, and can prove, be the matrix W of M * N for certain size _N, following formula is set up:

W _N(X _N)＝W _NX _N

To be defined as starting point in the equation (3), can obtain the variation W of any dimension N according to the recursive definition of DWCT _N(X _N) (N is 2 integer power), the situation when below N=4 being discussed.To a vector X ₄It is 2 vector that applicable equations (9) and (10) can obtain two length.The W that provides according to equation (3) ₂Definition and equation (5)-(8) can obtain Y ₄, according to equation (4), Y ₄Be to original vector X ₄Use the result behind the desirable DWCT.For general N=2 ⁿ, equation (9) and (10) constantly carry out the repetition valuation until obtain length be 22 ^N-1Till the individual vector.In fact, the recursive definition of equation (3)-(10) all needs to estimate when calculating the DWCT transformation matrix at every turn.In addition, transformation matrix W _NAlso can be according to equation (4) from original vector X _NBegin, constantly calculate and will be worth according to the N value and preserve to do next step calculating.The benefit of recursive definition is to be applied to any N value.The recursive definition of DWCT inverse transformation is expressed as IDWCT, because all equatioies all are linear, can be derived by equation (3)-(10).In addition, in some implementation procedure, inverse transformation also can directly be calculated and preserve from the forward direction conversion.

Array X _{N * M}(N=2 ⁿ, M=2 ^m) two-dimentional DWCT can be defined as:

Z _N×M＝{W _M{W _N(X _N×M)} ^T} ^T

At first the row of this two-dimensional array is done conversion, be about to each provisional capital and be used as one one n dimensional vector n; Then the row of array are done conversion, each row also is used as one one n dimensional vector n.Each colour band is DWCT respectively.Be that each chrominance component all is used as a monochromatic colour band and is extracted out, then each such colour band be DWCT respectively.

When the length of image is not 2 power,, will use a kind of low frequency expansion algorithm according to the present invention.At first consider one-dimensional case.If vector X _L={ x ₀, x ₁..., x _L-1} ^T, wherein, 2 ^N-1＜L＜2 ⁿ, will be increased 2 so thereafter ⁿ-L element x _L, x _L+1..., x _N-1, be extended to a new vector X _N, it is of a size of N=2 ⁿThe element of these new interpolations can be got arbitrary value, because they can be lost by decoded device at last.Present low frequency expansion is in order to select x _L, x _L+1..., x _N-1Value to X _NDo the DWCT conversion, in fact have only L nonzero coefficient useful.

Consider X _NThe DWCT inverse transformation: $X_N=\left[\begin{array}{c}{X}_L\\ X_{(N-L)}\end{array}\right]=W_N^{-1}\left(Z_N\right)=W_N^{-1}{Z}_N=\left[\begin{array}{cc}{W}_{\mathrm{LxL}}& W_{\mathrm{Lx}(N-L)}\\ W_{(N-L)\mathrm{xL}}& W_{(N-L)x(N-L)}\end{array}\right]\left[\begin{array}{c}{Z}_L\\ Z_{(N-L)}\end{array}\right]$

Z _(N-L)Be changed to 0, the result of following equation is:

X _L=w _LxLZ _L, perhaps Z _L=w _LxL ^-1X _L

X _(N-L)＝w _(N-L)xLZ _L＝w _(N-L)xLw _LxL ^-1X _L

In above-mentioned equation, X _(N-L)Be to initial data X _LThe expansion of low frequency.When two-dimensional case, consider the column vector of conversion image earlier, consider row vector (image is carried out transposition) again.Under two-dimensional case, establishing a block size is J * K, and J and K are not 2 integral multiples, and row and column all needs to add element so, and making has J * K nonzero value in the coefficient block after conversion at the most.Same rule also is applicable to other conversion such as DCT, Fourier transform or wavelet transformation.

In order to reduce the amount of calculation of DWCT, original picture should resolve into piece, and each piece is DWCT, rather than whole image is DWCT.We consider one-dimensional case earlier, at first select a piece size M=2 ^m, then given vector X _LBe decomposed into (L+M-1)/M piece.Only some is an initial data to possibility in last piece, will use above-mentioned low frequency extended method to come the remainder of filling block in this case.Expansion under the two-dimensional case can directly be divided into size to image and be the piece of M * M pixel.All colour bands all will carry out the piece operation splitting, so that all colour band images can both have identical overlapping block boundary.In order to obtain reasonable result, typical block size is 8 at least, though that the size of piece can be selected is bigger, in the practical operation of DWCT, 8 * 8 or 16 * 16 block size is only comparatively rational.And,, also can significantly increase the requirement of computational resource along with the increase of piece size.Traditional wavelet transformation uses bigger piece size through regular meeting, for obtain best quality in addition whole image as a bulk.The advantage of existing DWCT is not need to use maximum piece size, and keeps the advantage of wavelet transformation.

An image is divided into some process equally also can be applicable to 130 modules and selects other mapping algorithm for use.According to the flow chart of Fig. 1, image needs piecemeal before entering 130 modules, so that conversion module 130 and quantization modules 140 are that piece is operated.In the cyclic process of difference block 120, a width of cloth is being merged into an images by the image of piecemeal again through after the inverse transformation of module 160.

In quantization modules 140, each coefficient of piece all will quantize with quantized value, and with rounding as a result to immediate integer, thereby the fixed-point data of obtaining.Quantized value will be selected according to the requirement of using.The DWCT of the picture block of a M * M or other conversion meeting produce the coefficient block of a M * M.As mentioned above, piece also may be whole image.The set that a piece is done the quantized value of quantification has constituted a quantization parameter piece, is called quantization table.It in the table 17 structure of a quantize block.Table 17: quantization parameter

??Q 0，0	??Q 0，1	…		??Q 0，M-1
					??Q 1，0	??Q 1，1
??…
??Q M-1，0		…		??Q M-1，M-1

For example, establish Q (x y) is the quantization table of selecting, C (x is the DWCT coefficient of a picture block y), and then quantizing process can be expressed as follows:

D(x，y)＝[C(x，y)/Q(x，y)]

Wherein [X] expression is digital X rounding to an integer.

Quantization parameter Q (x, value y) has determined the storage precision of conversion coefficient, the highest precision is the data precision before the conversion.(figure place of each coefficient storage tails off for x, the raising of value y) along with the Q that selects.The conversion that easy rule is is N * N for a size, (x, minimum actual value y) is N to Q.If used the conversion of a nonstandardized technique, the first row Q of quantization table (0, also will adjust through corresponding, the influence that does not have standardization to bring with first row of compensation conversion by value y).In some cases, for the quantization table of a fixed number of a given image meeting use, so that a different coefficient block can be used a different quantization table.For with the coefficient block of highest frequency coefficient correspondence, need bigger quantized value, for example, can use the value about 500.

In inverse quantization module 150, the value in integer value and the quantization table multiplies each other.Inverse quantization module 150 reverse use quantization tables.

In sequencing statistical coding module 180, the quantization parameter of each piece is pressed following three steps can't harm form coding: context prediction 185, ordering 190 and entropy coding 195.

Traditional context Forecasting Methodology can be used in module 185.According to the present invention, a kind of improved context Forecasting Methodology is arranged, it has used a kind of each colour band and each coefficient context all inequality.For main color part P, context is included in the neighborhood coefficient in same.For first less important colour band S, context comprises neighborhood coefficient in same and the coefficient that corresponds to the same position on the main colour band.For second less important colour band Q, context is included in the neighborhood coefficient in same and corresponds to the coefficient with same position on preceding two colour band pieces.

Index value is that 0 coefficient has different Forecasting Methodologies, and its predicted value generally is that index value is 0 coefficient in the adjacent block, and the same position coefficient that is arranged in main colour band or less important colour band.The context of any exponent number can use, yet, generally all use second order or three rank.One three rank context is by three coefficient C ₁, C ₂, and C ₃Form, they are from same or different masses.A second order context only comprises C ₁And C ₂The context that describes below is three rank.The second order context can obtain by removing a coefficient, and this coefficient is the most weak one of correlation in the context normally.

Coefficient in piece can be divided into 4 groups, and every group context forms by different rules:

0 group: 0 group of coefficient that only comprises (0,0) position.

1 group: 1 group comprises first row except the coefficient outside the coefficient (0,0).This group coefficient can be expressed as: (0, i), i＞0.

2 groups: 2 groups comprise first row in except the coefficient outside the coefficient (0,0), be expressed as (j, 0), j＞0.

3 groups: 3 groups comprise all remaining coefficients, are expressed as: (i, j), i＞0, j＞0.

Main color coefficients P in 1 group _{0, i}The context of (i＞0) is by being positioned at delegation and at P _{0, i}3 coefficients before constitute:

C _k＝P _0，i-k，k＝1，2，3.

If C ₃Or C ₂Be in the piece outside, context is only by C so ₁Constitute.Same, a FACTOR P _{J, 0}Context in 2 groups by same list be positioned at P _{J, 0}3 coefficients before constitute.If C ₃Or C ₂In the outside of piece, context is also only by C so ₁Constitute.A FACTOR P _{J, i}At 3 groups contexts by being positioned at P _{J, i}Upper left 3 neighborhood coefficients constitute.

C ₁＝P _j，i-1，C ₂＝P _j-1，i，?C ₃＝P _j-1，i-1.

In a piece, the coefficient positions of 0 group, 1 group, 2 groups and 3 groups is represented by Fig. 3.

The context of coefficient (0,0) has different constructive methods.At first, all be integrated into from all coefficients (0,0) of different masses in the main colour band and constituted an index image together.This index image can resemble the piece of a normal main colour band image processed, this index image is used DWCT can obtain a coefficient block.1 group, 2 groups and 3 groups are predicted with above-mentioned context.For a image with H position precision, the coefficient of (0,0) position constant 2 ^H-1Prediction.

In 1 group, a coefficient S of less important colour band _{0, i}The context of (i＞0) is by with being positioned at S in the delegation _{0, i}Before 2 coefficients and corresponding main colour band FACTOR P _{0, i}Constitute.

C ₁＝P _0，i，C ₂＝S _0，i-1，C ₃＝S _0，i-2

If C ₃Be positioned at the piece outside, context is only by C so ₁And C ₂Constitute.Equally, 2 groups of interior coefficient S _{J, 0}Context by being positioned at S in the same row _{J, 0}Before two coefficients and corresponding main colour band FACTOR P _{0, i}Constitute.

C ₁＝P _j，0，C ₂＝S _0，i-1，C ₃＝S _0，i-2

If C ₃Be positioned at the piece outside, context only comprises C so ₁And C ₂Coefficient S in 3 groups _{J, i}Context by being positioned at S _{J, i}Upper left two neighborhood coefficients and corresponding main colour band coefficient constitute P _{J, i}

C ₁＝P _j，i，C ₂＝S _j，i-1，C ₃＝S _j-1，i，

The context of the coefficient (0,0) of the contextual formation of the coefficient of less important colour band S (0,0) and main colour band constitutes similar.At first, concentrate at from all coefficients (0,0) of less important each different masses of colour band and constitute an index image together.This an index image and common the same a processing of less important colour band piece are DWCT to this index image and are produced a coefficient block.Use above-mentioned context that 1 group, 2 groups and 3 groups are predicted for less important colour band.Precision is that the image of H uses a constant 2 ^H-1Come predictive coefficient (0,0).

The 3rd colour band coefficient Q in 1 group _{0, i}The context of (i＞0) is by with being positioned at S in the delegation _{0, i}Coefficient before and first corresponding colour band FACTOR P _{0, i}With second colour band coefficient S _{0, i}Constitute.

C ₁＝P _0，i，C ₂＝S _0，i，C ₃＝Q _0，i-1，

Same, the coefficient Q in 2 groups _{J, 0}Context by being positioned at Q in the same row _{J, 0}Coefficient before and first corresponding colour band FACTOR P _{J, 0}With second colour band coefficient S _{J, 0}Constitute.

C ₁＝P _j，0，C ₂＝S _j，0，C ₃＝Q _j-1，0，

Coefficient Q in 3 groups _{J, i}Context by being positioned at Q _{J, i}The neighborhood coefficient of left and first corresponding colour band FACTOR P _{J, i}With second colour band coefficient S _{J, i}Constitute.

C ₁＝P _j，i，C ₂＝S _j，i，C ₃＝Q _j，i-1，

The context of the contextual formation of the coefficient (0,0) of the 3rd colour band and (0,0) coefficient of above-described first colour band and second colour band forms similar.

Except the context by the neighborhood pixel, the position of each coefficient in the DWCT piece also can constitute a context.Location context is expressed as C ₀

The colour band of noting an image is in certain quality scale, for example under 4 grades pattra leaves sampling configuration, is not on same position.In this case, a pixel of certain color component can not find the pixel of a correspondence at the same position of other color component, and this may cause the color blocks disunity, even may cause the piece size inequality.The piece dimensional problem can be handled like this, promptly only uses the sub-piece of the main coefficient that is positioned at the upper left corner when producing context for less important color.Perhaps, ignore misalignment issues.In addition, context Forecasting Methodology described herein can be applied to any data compression method.For example, present context Forecasting Methodology can with JPEG or JPEG2000 etc. similarly the entropy coding module in the method combine use.

Pixel context and location context constitute a complete context for probabilistic model together.This complete context table is shown C ₃C ₂C ₁C ₀To some coefficient X, conditional probability P (C ₃C ₂C ₁C ₀| X) be used in 195 modules, coefficient X being encoded.In actual applications, for the consideration of computation complexity, contextual maximum order can be dropped to 3 or 2 from 4.

In order to carry out efficient coding, in order module 190, the pixel array of two dimension is mapped to an one-dimensional sequence.In the flow chart of Fig. 1, ordering is predicted followed by context.In addition, in case all related datas such as coefficient, quantization table and contextual mapping relations are saved, ordering just can have been operated before quantification and context prediction.Traditional sort method is included in the zig-zag ordering of using among the JPEG.The present invention has defined another kind of sort method: four fork numbers orderings, because DWCT is placed on the upper left corner of two-dimensional array to most important coefficient, so this sort method is specially adapted to the data after the conversion.At first, shown in table 18, each piece all is divided into 4 districts (PRs) of equal sizes, and each district has all been put on the sequencing of a numeral to represent that it is accessed.

Table 18: priority area

????0	????1
		????2	????3

In 4 data fields, indicate 0 zone and have the highest priority, indicate 3 zone and have minimum priority.In code stream, at first be the data in 0 district, secondly be the data in 1 district, then be only 2 district's data and 3 district's data.

Use above-mentioned same procedure, each zone can both further be divided into littler priority subregion.This process constantly repeats to reach 1 * 1 pixel size until area size.Shown in the table 19 that a size is the ranking results of 16 * 16 piece.Same method can be used for the piece of size arbitrarily.

Table 19: the quaternary tree ordering of piece coefficient

????0	????1	????4	????5	????16	????17	????20	????21	????64	????65	????68	????69	????80	????81	????84	????85
																????2	????3	????6	????7	????18	????19	????22	????23	????66	????67	????70	????71	????82	????83	????86	????87
????8	????9	????12	????13	????24	????25	????28	????29	????72	????73	????76	????77	????88	????89	????92	????93
																????10	????11	????14	????15	????26	????27	????30	????31	????74	????75	????78	????79	????90	????91	????94	????95
????32	????33	????36	????37	????48	????49	????52	????53	????96	????97	????100	????101	????112	????113	????116	????117
																????34	????35	????38	????39	????50	????51	????54	????55	????98	????99	????102	????103	????114	????115	????118	????119
????40	????41	????44	????45	????56	????57	????60	????61	????104	????105	????108	????109	????120	????121	????124	????125
																????42	????43	????46	????47	????58	????59	????62	????63	????106	????107	????110	????111	????122	????123	????126	????127
????128	????129	????132	????133	????144	????145	????148	????149	????192	????193	????196	????197	????208	????209	????212	????213
																????130	????131	????134	????135	????146	????147	????150	????151	????194	????195	????198	????199	????210	????211	????214	????215
????136	????137	????140	????141	????152	????153	????156	????157	????200	????201	????204	????205	????216	????217	????220	????221
																????138	????139	????142	????143	????154	????155	????158	????159	????202	????203	????206	????207	????218	????219	????222	????223
????160	????161	????164	????165	????176	????177	????180	????181	????224	????225	????228	????229	????240	????241	????244	????245
																????162	????163	????166	????167	????178	????179	????182	????183	????226	????227	????230	????231	????242	????243	????246	????247
????168	????169	????172	????173	????184	????185	????188	????189	????232	????233	????236	????237	????248	????249	????252	????253
																????170	????171	????174	????175	????186	????187	????190	????191	????234	????235	????238	????239	????250	????251	????254	????255

The output of order module comprises a table, and this table has shown coefficient order in the one-dimension array in the top diagram and the correlation between the two-dimentional pixel location.

The final step of data compression method is an entropy coding module 195.Based on the conditional probability that the context of determining produces, the coefficient through ordering, quantification is carried out entropy coding in 185 modules.Can adopt any lossless coding method, but preferentially use arithmetic coding.The arithmetic coding method of standard is at reference book TheData Compression Book (the M ﹠amp that is collaborateed by Mark Nelson and Jean-Loup Gailly; T Books, 1995) have a detailed description in.

The conditional probability of the coefficient X that encodes: P (C ₃C ₂C ₁C ₀| X) at first carry out initialization, upgrade with the accumulation probability statistics of particular image then by the constant table of a priori.Note C ₀It is location index by the one-dimension array of order module generation.If X is not at caluclate table C ₃C ₂C ₁C ₀In, then prediction order is kept to 3, caluclate table C ₂C ₁C ₀Be verified.If X is at caluclate table C ₂C ₁C ₀In, probability of use P (C so ₂C ₁C ₀| the X) X that encodes.Otherwise order reduces to 2, and check second order table C ₁C ₀This process lasts till that always exponent number is till 1.If X is not still at caluclate table C ₀In, be that all are not at caluclate table C so ₃C ₂C ₁C ₀, C ₂C ₁C ₀, C ₁C ₀And C ₀In uniform probability of value hypothesis, and X is encoded by this probability.

The value that need be encoded is called a character.The caluclate table of different rank is incoherent, and they do not comprise common character.For example, if X not at caluclate table C ₃C ₂C ₁C ₀In, after we drop to 3 with exponent number from 4 so, when calculating its character probabilities, all C ₃C ₂C ₁C ₀In character should be from caluclate table C ₂C ₁C ₀In remove.Equally, if we drop to 2 (because X is not at caluclate table C with exponent number from 3 ₂C ₁C ₀In), when calculating its character probabilities, all are contained in C ₃C ₂C ₁C ₀And C ₂C ₁C ₀Character all should be from caluclate table C ₁C ₀In remove.In cataloged procedure, each conditional probability is all represented with occurrence number.During certain value of each coding, the occurrence number of corresponding this value all will be updated.

Coding module is a process that produces code stream, and the code stream behind the coding can effectively be stored or effectively transmit on computer network.Code stream comprises it at first being code coefficient corresponding to the gross grade, then is the code coefficient corresponding to difference image.In addition, code stream may comprise general header, the arrangement information of file size, color number, number of samples and this file for example is as the sign of a piece sequence index, using method, as the quantization table of conversion and coding, appointment and optional probability tables etc.

As shown in Figure 2, a decompression process is that each step in the compression process is carried out inverse operation basically, and image reconstruction in the end.Decompression process needs to visit and is used for the probability tables of entropy coding.Though probability tables can be used as the part storage of header, but they are generally all very big.So the probabilistic forecasting table is produced by a similar context forecasting process that uses in the module 185 in module 210.Yet the context forecasting process has used the information of neighborhood pixel.The part of the decompression process of circulation 280 expressions in different pixel location cyclings, is progressively built probability tables, and circulation 280 is inverse process of sequencing statistical coding.At the entropy decoder module, code stream is resumed into the conversion coefficient after the quantification.Contrary order module 230 is shone upon back two-dimensional array to one dimension coefficient array.Inverse quantization module is used a converse quantization table reduction quantized data.Inverse transform module 260 becomes pixel value to the data conversion behind the inverse quantization.Notice that module 250 and 260 is just in time used and the circulation of difference block 120 in module 150 and 160 in the identical operations process.

As shown in Figure 2, reconsolidate the influence that process 290 has been removed difference block 120.A difference image sequence of inverse transform module 260 outputs is from gross grade image G _SBeginning is expressed as { G _S, G _S-1... G _C, owing to experienced the cause that diminishes conversion, G ' s herein and original picture I _SAnd it is incomplete same.G _SEnlarge and produce a more high-order image H _S-1The F that rebuilds _S-1, be to I _S-1Approximate, according to equation (2) F _S-1=G _S-1+ H _S-1Obtain.Image F _S-1Be extended to next stage image: H _S-2, this process constantly repeats to be resumed out until the data of first water grade image.According to normal definitions H _S=0, the merging process of describing in the flow chart as Fig. 2 290, wherein, shown in equation (2), addition module 265 is with difference image G that rebuilds and the visual H addition that obtains from lower grade expansion.

Yet, not necessarily leave no choice but the image reconstruction that is stored in the first water grade in the packed data is come out.If by the display quality grade (DQS) of applied environment appointment, be a kind of credit rating also lower than the credit rating of Image storage, need only the difference of recovering corresponding to the DQS grade so.In this case, the requirements of process among Fig. 2 changes, so that can once obtain from the difference image G of module 210 to 260, and before the difference image that is positioned at next credit rating carries out inverse transformation, be added on the expansion image of front.Under this mode, before more the information that requires of high-quality level was calculated, all images that are used for showing on a given credit rating will be determined at any next one.

Therefore, according to above step, the image of a compression can be watched, download or transmit in progressive mode by computer network or internet.More have, browser can carry out image in the credit rating of an appointment and show, and need not comprehend corresponding to any data of high-quality level more.

Though the digital image compression process is described by sampling definition, conversion and the ordered steps of appointment, this only describes a just example of the present patent application.For the various distortion of the technology of the present invention with revise and all will think the category that belongs to the present invention's invention.

The inventive method can realize by software, specialized hardware or soft, combination of hardware.

Claims (10)

1. the compression method of a digital image is compressed into an one dimension code stream to the numeral of an image, and the numeral of image comprises a two-dimentional pixel array, and wherein each pixel all interrelates with a main colour band and less important colour band, it is characterized in that:

Image table is shown as the credit rating that a series of quality is successively decreased gradually, wherein, the data that the better quality grade comprises relatively low-quality level want many, reduce the colour band number of higher quality grade or the number of minimizing higher quality grade pixel and can obtain lower credit rating;

Represent an image according to each credit rating, described image comprises a gross grade image and difference image;

The difference image of some credit ratings is to do difference by two images to obtain, and an image is the realistic images of this grade, and another image is to obtain amplifying at the image reconstruction than low-quality level;

By a process gross grade image and difference image are expressed as integer value, this process comprises: earlier image transform is become the one group of coefficient that interrelates with known function, with quantized value this group coefficient sets is quantized then and rounding arrives integer value;

Coding makes integer-valued numeral produce the code stream of an one dimension corresponding to the integer value of minimum quality grade and difference image with harmless sequencing statistical coding method.

2. according to the compression method of the described digital image of claim 1, it is characterized in that: the method for determining image reconstruction comprises: the coefficient that quantized and quantized value are multiplied each other carry out inverse quantization; The inverse transformation of carrying out interrelating with known function produces one and rebuilds expression.

3. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: when being in rgb space, main colour component is meant green, and less important colour component is meant redness and blueness; When being in yuv space, main colour component is meant brightness, and less important colour component is meant colourity.

4. according to the compression method of the described digital image of claim 1, it is characterized in that: the credit rating sequence comprises: first credit rating, and wherein each pixel all comprises all colour components; Second credit rating, wherein each pixel all comprises main colour component and a less important colour component; The 3rd credit rating, wherein each pixel all comprises a main colour component, and its quantity is 4 times of each less important colour component quantity; The 4th credit rating, wherein each pixel all only comprises a colour component, and the quantity of main colour component is the twice of less important colour component quantity; The 5th credit rating is derived from first credit rating, by both direction all divided by an integer ratio factor, reduce the number of pixels of its horizontal direction and vertical direction; The 6th credit rating is derived from the 4th credit rating, and wherein each pixel comprises main colour component and a less important colour component; The 7th credit rating derives from the 4th credit rating, and wherein each pixel all comprises a main colour component, and its quantity is 4 times of each less important colour component quantity.

5. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: the method that obtains one group of coefficient interrelating with known function is to use a two-dimensional transform that combines frequency domain character and many resolutions feature, and the process of two-dimensional transform comprises:

Before doing conversion, the data of minimum quality grade and difference image are resolved into the square of M * M, M is 2 integer power;

If the piece that size is J * K, J and K are not 2 integer powers, at first will mend into it the square of a M * M, can realize so that after the conversion, J * K nonzero value being arranged at the most in the coefficient sets by replenish new element in the original block of J * K;

Two-dimensional transform comprises the one-dimensional transform of two-dimensional array being done a line direction, again this array is done an one-dimensional transform on the column direction;

Be based on the one-dimensional transform of using discrete cosine transform and permutation function adopt the recursive fashion definition, and the output element of conversion can be divided into even number element and odd elements, and the even number element comprises the expression than low-quality level in the input data of conversion.

6. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: harmless sequencing statistical coding method comprises context prediction, ordering and the entropy coding of quantization parameter, wherein:

The context prediction is operated respectively according to each colour band: for main colour component, context comprises the adjacent coefficient of a location index and main color pixel; For first less important colour component, context comprises adjacent coefficient in a location index, this colour band and the coefficient that is positioned at the main colour band of same position; For second less important colour component, context comprises a location index, the adjacent coefficient in this colour band and be positioned at main colour band coefficient and first less important colour band coefficient on the same position;

The context prediction is divided into four groups with quantization parameter: first group of zero coefficient that comprises corresponding to the extreme lower position index; Second group comprises all data of removing the row of first outside first group of data; The 3rd group comprises all first columns certificates of removing outside first group of data; The 4th group comprises all residual coefficients.The context of each group is all different;

Sorting operation is that two-dimensional array is arranged in one-dimension array, arrangement comprises a two-dimensional array is divided into 4 equal and opposite in direction zones, and to set access order be upper left, upper right, lower-left and bottom right, and constantly subregion is repeated till area size is 1 * 1 pixel in each zone;

The entropy coding process is that the numeral of image is compressed into a code stream, numeral comprises the pixel array of a two dimension, wherein each pixel all comprises a main colour component and less important colour component, and the probability tables that will use in the entropy coding process is formed by the context Forecasting Methodology.

7. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: the decompression method of digital image comprises the steps:

Recover the context prediction probability table of use in the compression;

Code stream is decoded, recover integer value corresponding to gross grade image and difference image;

The adjustment order is arranged in a two-dimensional array form to decoded integer;

Each desorption coefficient corresponding to gross grade and difference image is multiplied each other with quantized value;

Do the inverse transformation that interrelates with known function and rebuild the numeral of gross grade and difference image;

The image that is positioned at than low-quality level is extended to next higher quality grade;

A given credit rating, difference image and the visual addition that is extended to this grade are obtained the numeral of a reconstruction of this grade.

8. according to the decompression method of the described digital image of claim 7, it is characterized in that: with inverse transformation that known function interrelates is the conversion that frequency domain transform characteristics and many resolutions transform characteristics are combined, this conversion is to carry out recursive definition on the basis of an inverse discrete cosine transformation and a permutation function, wherein, the input data of inverse transformation have been divided into even number element and odd elements, and the even number element comprises the more low-level expression in the input data of this conversion.

9. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: credit rating realizes by the sub-sampling sequence, in a sample sequence, the length of sampled data is successively decreased, sampled data represents to comprise the pixel array of a two dimension, wherein, each pixel all has main colour component and less important colour component, and this sequence comprises following particular sequence:

First kind of sampled representation, each pixel all has all colour components;

Second kind of sampled representation, each pixel all have main colour component and a less important colour component;

The third sampled representation, each pixel all have a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity;

The 4th kind of sampled representation comes from first kind of sampled representation, but the dimension of its two-dimensional array is divided by an integer ratio factor, the corresponding in the horizontal and vertical directions number that reduces pixel;

The 5th kind of method of sampling is derived from the 4th kind of sampled representation, and each pixel all comprises main colour component and a less important colour component;

The 6th kind of sampled representation is derived from the 4th kind of sampled representation, and each pixel all comprises a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity.

10. according to the compression method of claim 1 or 2 described digital images, it is characterized in that: credit rating realizes by the sub-sampling sequence, in a sample sequence, the length of sampled data is successively decreased, sampled data represents to comprise the pixel array of a two dimension, wherein, each pixel all has main colour component and less important colour component, and this sequence comprises following particular sequence:

First kind of sampled representation, each pixel all has all colour components;

Second kind of sampled representation, each pixel all have main colour component and a less important colour component;

The third sampled representation, each pixel all have a main colour component, and the quantity of main colour component is 4 times of each less important colour component quantity;

The 4th kind of sampled representation, each pixel comprises a color component, and the quantity of main color component is the twice of less important color component quantity.

Images